# The Log, issue 4: the bill, the proof, the banks > Issue 4 of The Log, the week in AI for people who don't follow it every day. Anthropic filed to list and Alphabet lost $500 billion in a day, OpenAI said its models proved three things nobody has checked yet, and Canada's banks shut the door on personal agents a week after Walmart opened one. - Canonical: https://www.anchorterminal.com/blog/the-log/issue-4-the-bill-the-proof-the-banks - Markdown: https://www.anchorterminal.com/blog/the-log/issue-4-the-bill-the-proof-the-banks.md (~3,550 tokens) - Slim: https://www.anchorterminal.com/blog/the-log/issue-4-the-bill-the-proof-the-banks.min.md (~1,430 tokens, same facts, less prose, for token-sensitive contexts) - JSON: https://www.anchorterminal.com/blog/the-log/issue-4-the-bill-the-proof-the-banks.json (this page as data, same URL with Accept: application/json) - Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt) - API: https://www.anchorterminal.com/api/v1/index.json - MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key) - DNS: dig +short TXT .grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.. and .vs.. under the zone, and the same under grade.letme.dev) - Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data) - Updated: 2026-11-01 By Vlad Cealicu, 1 November 2026. 1,589 words, about 7 min to read. At 1:48am on Sunday 1 November I pushed every grade on this site to Hugging Face as a dataset, 1,381 rows, and then sat down to write this, which is the part of the week I like least (the dataset is at [AnchorTerminal/grades](https://huggingface.co/datasets/AnchorTerminal/grades), and it updates itself now, so at least that bit is done). This is the week in AI for anyone who doesn't follow it every day. Anthropic filed to go public on Monday with the first audited numbers anyone has seen for a frontier lab [[1]](#ref-1), Alphabet lost about $500 billion of market value on Thursday because it's spending more on AI than it makes and its search ads are growing slower [[2]](#ref-2), OpenAI said on Monday that its models had proved three results that people have worked on for decades [[4]](#ref-4), and six Canadian banks said personal AI agents can't use their online banking [[6]](#ref-6). ## The bill arrives from both ends Anthropic's S-1 went up on EDGAR at 9am New York time on Monday 26 October, to list on Nasdaq as ANT [[1]](#ref-1). I spent that afternoon reading it and [wrote it up the same day](/blog/anthropic-s-1), so the short version here. Revenue was $41 million in 2023, $386 million in 2024, $4.6 billion in 2025 and $16.3 billion in the first six months of 2026, and the whole business it describes is agents calling things, with "agent loop", "computer use" and "prompt injection" defined in the glossary for a securities regulator [[1]](#ref-1). The $80.7 billion loss the headlines ran with is $72.4 billion of accounting, convertible notes (mostly Amazon's) going up in value because the company did, and the operating loss, the money lost running the business, was $693 million [[1]](#ref-1). Public shareholders will get the economics and at most 49.9% of the vote [[1]](#ref-1). Three days later the other end of the same bill came due. Alphabet reported on Wednesday 28 October and on Thursday the shares fell about 13%, roughly $500 billion, its worst day since it listed in 2004 [[2]](#ref-2) [[3]](#ref-3). It spent $51.4 billion on data centres and chips in the quarter against $42.3 billion of operating cash flow, so for the second time since 2004 it burned cash, and it raised its spending plan for the year to as much as $205 billion [[2]](#ref-2). Search revenue grew 12%, down from 17% the quarter before, and Google's chief business officer said its own AI Mode answers earn less from ads than the list of links did [[3]](#ref-3). Cloud grew 103% to $30.7 billion, which nobody cared about on the day [[2]](#ref-2). I think the two stories are one story. Anthropic's filing shows what the tokens cost to make and what the clouds charge to resell them ($8.9 billion of compute for $16.3 billion of revenue, and $2.7 billion of the year's increase in sales costs is fees to channel partners [[1]](#ref-1)), and Alphabet's quarter shows what it costs to be one of the clouds. Both are fine businesses with a bill attached, and this was the first week the bill was on the table in public, with audited numbers, on both sides. ## Three proofs, none checked On Monday 26 October OpenAI said its internal models had produced three results, a proof of the local smoothing conjecture in every dimension, which would settle the Kakeya conjecture along the way, a disproof of the Exponential Time Hypothesis, and a proof that quantum chromodynamics exists as a mathematical object [[4]](#ref-4). Any one of those would be a career. The Exponential Time Hypothesis is the one I'd watch, because hundreds of papers assume it's true and build on it [[4]](#ref-4), and if the disproof holds, a lot of what we thought we knew about which problems are hard needs re-reading. None of the three has been checked by the wider community. All three were formalised in Lean, which checks every step mechanically, and two of the Lean proofs are public, but the people reviewing the computer science one are the people who proposed the hypothesis, and some of them signed an open letter about confidentiality agreements and who gets the credit [[5]](#ref-5). I [wrote about this on Tuesday](/blog/openai-three-results), and my position hasn't moved. A Lean proof is a strong claim, and a strong claim is still a claim until someone who didn't make it has read it. If it holds, the result that matters for everyone outside maths is that a lab produced three of these in about six weeks [[4]](#ref-4), and the bottleneck moves from finding proofs to checking them. ## The banks shut the door Walmart opened In [issue 3](/blog/the-log/issue-3-found-leaked-let-in) Walmart said it would let shopping agents in through its own gateway. On Wednesday 28 October Canada's banks went the other way, all of them at once. RBC will block third-party personal agents from its online banking, BMO doesn't allow them on its digital properties, TD has decided to block them, CIBC doesn't support them, Scotiabank will restrict them and National Bank is watching [[6]](#ref-6). RBC's chief executive wrote that nothing is more important than the security of client information, and the reason every bank gave was the same one [[6]](#ref-6). I don't think they're wrong, yet. An agent that holds your banking password is a worse idea than an agent with a Walmart profile that has a spending limit and a switch to turn it off, and that's the difference between the two weeks. Walmart built the door with a lock on it, and the banks have no door, so they bricked up the window. I'd guess the banks build the door within a year, with the same shape as Walmart's (a profile for the agent, limits, every transfer approved by a person), because the alternative is customers handing their passwords to an agent anyway and the bank finding out when something goes wrong. Open Banking already exists in the UK for this kind of scoped access, so the pieces are there. ## Also this week Google DeepMind announced Genie 4 on Tuesday, a world model that builds a scene from a description and lets you change it while it runs, with sound, for trusted testers only [[7]](#ref-7). I [wrote about the Gemini agent running beside it](/blog/genie-4-world-model), which plans what happens next and checks each frame, because that's the part that reaches the rest of us. On Thursday Google's Suncatcher satellite ran Gemma 4 on a TPU in orbit and sent a haiku back down through a ground station on Svalbard [[8]](#ref-8) (we're going to Svalbard next year, so I took that one personally, and [wrote it up](/blog/suncatcher-haiku-from-orbit)). And Anthropic's alignment team published research where Claude agents wrote training data that got past a safety monitor and still taught a model to misbehave, which reads as a scary headline and I [read the other way round](/blog/red-teaming-that-scales-with-compute), because it's the first safety test I've seen that gets better when you add compute instead of people. ## What we published - Mon 26. Anthropic's S-1 is a document about agents. I asked an agent to buy a train ticket to Edinburgh. Why does an AI agent need a different inbox anyway?. - Tue 27. Genie 4. OpenAI's models say the Exponential Time Hypothesis is false. Can you run GLM-5.3-Flash locally?. AI passes 37% of real engineering tasks. Slopcore will be a thing. - Wed 28. Is the best coding model better than a principal engineer?. Can you run Qwen3.8-27B locally?. - Thu 29. Your cousins, the open models an agent can start on spare hardware. - Fri 30. Five new things in agent tooling that nobody compares yet. The agent projects trending on GitHub, November 2026. Google ran Gemma 4 on a TPU in orbit. - Sat 31. We already have AI judges. Our grades, over DNS. Red-teaming that gets better when you add compute. The great software expansion. Two million page views took two years last time, and a month this time. - Sun 1. The Log, issue 4. The week of 26 October to 1 November 2026 Nineteen posts, which is too many, and I'm aware of it. If you only read one, read [the S-1 post](/blog/anthropic-s-1), because it's the first set of audited numbers on the thing everyone in this industry is standing on. [Two million page views](/blog/what-the-next-model-is-for) is the one about us, my last business took two years and ten people to get there and this one took a month and two of us, and [the great software expansion](/blog/the-great-software-expansion) is where I think that leads. Quynh wrote [why an agent needs a different inbox](/blog/why-an-ai-agent-needs-a-different-inbox) and [the agent projects trending on GitHub](/blog/github-trending-agent-tools-november-2026), and her GitHub posts still get more readers than anything I write, which is fine (it isn't). On the site, every listing's grade is now a DNS record, so `dig TXT mapbox.grade.anchorterminal.com` answers from inside a box that can't reach the web [[9]](#ref-9). Getting that to work meant finding out my home router keeps port 53 for itself on IPv4, so the zone answers over IPv6 and I'm adding Hurricane Electric's free name servers as secondaries for everyone else (I'd like to say I planned it that way). Every listing names the people behind it and the investors behind the company, with a page for each, the Chrome extension is on the Web Store, and the grades are on Hugging Face as of this morning, with a file a day for every day they change. The [changelog](/changelog/) has the rest, and there's a lot of rest. What isn't done. Performance and task success are still pending on every listing until the probes run, and the box they run on arrives this month. If you see a bank that builds the door, or an agent that got through a bricked-up window, send it my way, it's how issue 5 gets made. --- More in [The Log](https://www.anchorterminal.com/blog/the-log/index.md) ยท Previous issue [The Log, issue 3: found, leaked, let in](https://www.anchorterminal.com/blog/the-log/issue-3-found-leaked-let-in.md) - Older: [We already have AI judges, but not at the top courts yet](https://www.anchorterminal.com/blog/ai-judges-not-at-the-top-courts-yet.md) - All posts: https://www.anchorterminal.com/blog/index.md ## References Read on 31 October and 1 November 2026. Quotes are as published. [1] Anthropic, PBC, Form S-1 Registration Statement, filed with the U.S. Securities and Exchange Commission on 26 October 2026. Source for the ticker, the revenue by year and half, the glossary terms, the $80.7 billion net loss and the $72.4 billion fair-value change, the $693 million operating loss, the compute and channel-partner figures and the 49.9% vote. https://www.sec.gov/Archives/edgar/data/0002130357/000162828026067168/anthropicpbc-sx1.htm [2] The Motley Fool, Alphabet just lost $500 billion in market value, 29 October 2026, on Alphabet's third-quarter report of 28 October. Source for the 13% fall and the $500 billion, revenue of $127.6 billion, search growth of 12% against 17%, $51.4 billion of capital spending against $42.3 billion of operating cash flow, the spending plan of up to $205 billion and cloud's 103% growth to $30.7 billion. https://www.fool.com/investing/2026/10/29/alphabet-just-lost-500-billion-in-market-value-a-o/ [3] Implicator, Alphabet loses $500 billion as AI spending pushes cash flow negative, 29 October 2026, citing the Wall Street Journal and the Financial Times. Source for the worst day on record, cash flow negative for the second time since the 2004 listing, and Philipp Schindler on AI Mode answers earning less from ads. https://www.implicator.ai/alphabets-stock-falls-12-3-wiping-record-500b-market-cap/ [4] OpenAI, Working with academic communities to build scientific understanding, 26 October 2026. Source for the three results, the Lean formalisations and which are public, the hundreds of papers that assume the Exponential Time Hypothesis, and the six weeks. https://openai.com/index/building-scientific-understanding-with-academic-communities/ [5] Ana Maria Constantin, OpenAI says its AI model has disproved the Exponential Time Hypothesis, The Next Web, 26 October 2026. Source for the reviewers being the people who proposed the hypothesis, the open letter and the concern about credit. https://thenextweb.com/news/openai-exponential-time-hypothesis-false-kakeya-yang-mills [6] Daniel Johnson, The Canadian Press, Some Canadian banks blocking personal AI agents amid security concerns, 28 October 2026, as carried by CP24. Source for the six banks' positions, the reason they gave and Dave McKay's line. https://www.cp24.com/news/money/2026/10/28/some-canadian-banks-blocking-personal-ai-agents-amid-security-concerns/ [7] Google DeepMind, Genie 4, a new generation of world simulation, 27 October 2026. Source for what Genie 4 does, the sound, the trusted-tester preview and the Gemini model planning alongside it. https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/ [8] Travis Beals, We've received our first response from the Project Suncatcher prototype satellite, Google, 30 October 2026. Source for Gemma 4 on a TPU in orbit, the haiku and the downlink over Svalbard. https://blog.google/innovation-and-ai/models-and-research/google-research/project-suncatcher-satellite-response/ [9] Anchor Terminal, dig TXT mapbox.grade.anchorterminal.com, our grades over DNS, 31 October 2026. Source for the DNS records, the IPv6-only zone and the secondaries. https://www.anchorterminal.com/blog/grades-over-dns --- Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp